Daily electricity consumption determination method based on dynamic load identification

Abstract

The invention discloses a daily electricity consumption determination method based on dynamic load identification. The daily electricity consumption determination method comprises the following steps: step 1, high-frequency feature collection: obtaining a feature vector; step 2, loading a feature library: calling a reference feature matrix Fref, and matching the feature vector Fk and the reference feature matrix Fref by using a window constraint type dynamic time warping (DTW) algorithm; and step 3, determining daily electricity consumption Eday. According to the daily electricity consumption determination method based on dynamic load identification, an equipment feature library does not need to be marked manually, load fingerprints are automatically constructed through PCA and dynamic clustering, and the marking cost is reduced by 100%. The feature vectors are used for matching through the improved window constraint DTW algorithm, the matching delay is low, the real-time requirement is met, the energy loss of transient processes such as equipment starting and stopping and state switching is quantified, and the daily electricity consumption accounting precision is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of smart meter technology and relates to a method for determining daily electricity consumption based on dynamic load identification. Background Technology

[0002] The existing technology for determining daily electricity consumption has the following problems:

[0003] 1. Ignoring transient losses: Traditional methods only calculate steady-state power consumption during power consumption, resulting in a calculation error of up to ±15% for transient process losses such as air conditioner start-up and shutdown and motor switching when decomposing power consumption.

[0004] 2. Confusion of composite loads: When multiple devices are running in parallel, it is difficult to distinguish superimposed loads based on a single characteristic (such as power factor).

[0005] 3. High data dependence: Requires manual annotation of equipment feature database.

[0006] Therefore, a new method for determining daily electricity consumption based on dynamic load identification is needed to solve the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method for determining daily electricity consumption based on dynamic load identification, so as to solve the problems of the prior art.

[0008] The technical solution of the present invention is as follows:

[0009] A method for determining daily electricity consumption based on dynamic load identification includes the following steps:

[0010] Step 1: High-frequency feature acquisition: Acquire the current waveform I(t) and voltage waveform V(t) at a sampling rate α, where α ≥ 2kHz; extract the feature vector F using the obtained current waveform I(t) and voltage waveform V(t). k The feature vector F k Including harmonic energy distribution, VI trajectory curvature, and current change rate.

[0011] Step 2, Load the feature library: Call the reference feature matrix F ref =[F1,F2…F m The window-constrained dynamic time warping (DTW) algorithm is used to analyze the eigenvector F. k and reference characteristic matrix F ref The matching process is performed to obtain the matching results, which include the device ID and the device startup time t. start and equipment downtime t end, where the window constraint is: path offset radius: r = [0.2 × max(n, m)], where n is the length of the real-time monitoring sequence and m is the length of the reference template sequence;

[0012] Path offset range: |ij|≤r, where i is the i-th sampling point in the real-time monitoring sequence and j is the j-th sampling point in the reference template sequence;

[0013] Multi-feature weighting: The distance is calculated using the following formula:

[0014]

[0015] In the formula, d ω (i,j) is the weighted composite distance of similarity between the i-th point in the real-time monitoring sequence and the j-th point in the reference template sequence, I i To monitor the effective current value at the i-th point in the sequence in real time, I j K is the effective value of the current at the j-th point in the reference template sequence. i To monitor the VI trajectory curvature at the i-th point in the sequence in real time, K j For the curvature of the VI trajectory at the j-th point in the reference template sequence, Let be the rate of change of current at the i-th point in the real-time sequence. The current change rate at the j-th point in the reference template sequence;

[0016] Step 3: Determine the daily electricity consumption E using the following formula. day :

[0017]

[0018] In the formula, E base For base charge, ΔE k For the transient compensation power of the k-th device, Let k be the startup time of the kth device. Let k be the downtime of the device. Let α be the rate of change of current for the k-th device. k Let P be the compensation coefficient for the k-th device. k (t) represents the power of the k-th device at time t.

[0019] Furthermore, the base load E base Real-time estimation via Kalman filtering:

[0020]

[0021] In the formula, P base Δt represents the base load power, and Δt represents the base load operating time.

[0022] Furthermore, the curvature of the VI trajectory in step 1 is expressed by the following formula:

[0023]

[0024] In the formula, κ is the curvature of the VI trajectory, V is the voltage value, and I is the current value. The first derivative of voltage with respect to time. This is the second derivative of voltage with respect to time. This is the first derivative of the current with respect to time. It is the second derivative of the current with respect to time.

[0025] Furthermore, step 2 also includes a transient process identification step, utilizing the device startup time t. start and equipment downtime t end The transient time period of the k-th device is marked as Δt. k =[t start -0.5s, t end +1s].

[0026] Furthermore, when matching fails in step 2, principal component analysis (PCA) is used to decouple features and separate composite load features.

[0027] C = F T ·W

[0028] In the formula, C is the p×n principal component score matrix, W is the m×p loading matrix, F is the n×m original feature matrix, and F T Let F be the transpose of F.

[0029] Furthermore, the compensation coefficient α of the k-th device k Dynamic adjustment based on device type includes the following steps:

[0030] Step S1: Initial Matching: After identifying the type of the kth device, retrieve the initial value α of the calibration compensation coefficient corresponding to that device type from the pre-stored device feature library. preset ;

[0031] Step S2: Online Learning and Adaptive Correction: During the subsequent operation of the device, the actual energy consumption during its start-up and shutdown processes is continuously monitored. Based on multiple monitoring data, the compensation coefficient is adjusted using the following iterative algorithm:

[0032] Step S21: Calculate using the current coefficients Predicted transient loss E pred Compared with the actual back-calculated loss E actual The relative deviation δ between them;

[0033] Step S22: If the absolute value of the relative deviation average value If the threshold is exceeded, the compensation coefficient is adjusted according to the following formula.

[0034]

[0035] In the formula, μ is the learning rate. This represents the result of the t-th adjustment of the compensation coefficient. The result of the (t+1)th adjustment of the compensation coefficient is given, and δ is the predicted transient loss E. pred Compared with the actual back-calculated loss E actual The relative deviation between them;

[0036] Step S23: Update the It is written back to the device's feature record for subsequent calculations.

[0037] Furthermore, the compensation coefficient for the air conditioner is 1.12, and the compensation system for the motor is 1.05.

[0038] Furthermore, step 3 also includes an anomaly detection step, calculating the daily electricity consumption E obtained in step 3. day and total meter reading E meter The deviation rate δ, when the deviation rate δ≤5%, then the daily electricity consumption E is output. day Otherwise, repeat steps 2 and 3.

[0039] Beneficial Effects: The daily electricity consumption determination method based on dynamic load identification of this invention eliminates the need for manual labeling of equipment feature libraries. It automatically constructs load fingerprints through PCA and dynamic clustering, significantly reducing labeling costs. The improved window-constrained DTW algorithm utilizes feature vector matching, resulting in low matching latency and meeting real-time requirements. It quantifies energy losses during transient processes such as equipment start-up, shutdown, and state switching, compressing the transient error of ±15% in traditional methods to within ±3.2%, significantly improving the accuracy of daily electricity consumption calculation. Attached Figure Description

[0040] Figure 1 Flowchart for calculating the curvature of the VI trajectory;

[0041] Figure 2 A schematic diagram illustrating the real-time monitoring of the matching between the sequence and the reference template sequence;

[0042] Figure 3 A flowchart for the improved DTW algorithm;

[0043] Figure 4 Flowchart for decoupling PCA features;

[0044] Figure 5 This is a flowchart for dynamically adjusting the compensation coefficient based on the equipment type. Detailed Implementation

[0045] The method for determining daily electricity consumption based on dynamic load identification of the present invention includes the following steps:

[0046] 1. High-frequency feature acquisition:

[0047] Acquire the current waveform I(t) and voltage waveform V(t) at a sampling rate of ≥2kHz;

[0048] Synchronous extraction:

[0049] Harmonic energy distribution (3rd-21st harmonics);

[0050] Please see Figure 1 As shown, the curvature of the VI trajectory:

[0051] In the formula, κ represents the curvature of the VI trajectory, the rate of change of the load's dynamic admittance. A larger κ value indicates a more pronounced curvature of the trajectory in the (V,I) plane, and stronger nonlinearity, time-varying characteristics, or energy storage properties of the load. This parameter is the "DNA fingerprint" of the load type. For example: Purely resistive load: The VI trajectory is a straight line, κ≈0. Motor / inductive load: The VI trajectory is an ellipse, and the κ value changes periodically. Switching power supply: The VI trajectory is an irregular polygon, with a high and abrupt κ value. V represents the voltage value, and I represents the current value, the grid voltage (volts) and load current (amperes) at a given sampling time. The first derivative of voltage with respect to time. This is the second derivative of voltage with respect to time. This is the first derivative of the current with respect to time. The second derivative of the current with respect to time; the rate of change of the current. (Capturing transient features);

[0052] 2. Dynamic load matching

[0053] The window-constrained dynamic time warping (DTW) algorithm is adopted:

[0054] Window constraint: Path offset radius: r = [0.2 × max(n, m)], where n is the length of the real-time monitoring sequence and m is the length of the reference template sequence. The coefficient 0.2 is based on empirical data calibrated through numerous electrical equipment experiments. In power loads, the transient processes of equipment start-up and shutdown, state switching, and steady-state minor distortions caused by voltage fluctuations typically do not exceed 20% of the entire operating cycle. This constraint effectively filters out path distortions caused by noise or irrelevant fluctuations, while ensuring that true state changes are aligned with the path, reducing the computational complexity from O(n × m) to approximately O(0.2 × max(n, m)). 2This achieves a balance between accuracy and speed.

[0055] Suppose you are identifying an air conditioner.

[0056] Reference template m: The standard startup-run fingerprint of the air conditioner in the feature library, consisting of m = 800 data points (e.g., approximately 0.167 seconds at a sampling rate of 4.8 kHz).

[0057] Real-time sequence n: A segment of unknown load characteristics currently collected, with a length of n = 1000 points.

[0058] Calculate the window radius r:

[0059] r=[0.2×max(1000,800)]=200.

[0060] Constraints during matching:

[0061] When calculating the DTW cumulative distance matrix, the algorithm only allows searching within a strip region of 200 points on each side of the main diagonal of the path (i.e., constraint |ij|≤200). This forces that the i-th point in the real-time sequence can only be matched with a point in the reference template sequence whose index j is in the range [i-200, i+200].

[0062] Please see Figure 2 As shown, the path offset range is limited to: |ij|≤r, r=50~200, which is positively correlated with the sampling rate;

[0063] Where i is the i-th sampling point in the real-time monitoring sequence X. It represents the i-th sampling point of the high-frequency characteristic data (such as current, harmonics, VI trajectory) being collected from the total power line. It represents the actual power consumption status at the current time. j is the j-th sampling point in the reference template sequence Y. It represents the j-th sampling point of the standard feature fingerprint of a known electrical appliance (such as an air conditioner) in the feature library. It represents the standard state of the appliance in a typical operating phase.

[0064] During matching, the i-th point of the real-time detection sequence can only be compared and its path extended with the points in the reference template sequence whose index j falls within the window range of [ir, i+r].

[0065] For electrical loads, equipment operation is a causal and relatively stable physical process in terms of time. The startup surge of an air conditioner cannot be aligned with the steady-state operation 10 seconds later. Therefore, the constraint |ij|≤r reflects the prior knowledge that "the real-time state can only be matched with the reference state that is temporally adjacent," and the forced alignment path conforms to physical laws.

[0066] Suppose we are matching a real-time data segment (n=1000 points) using an air conditioner startup template (m=800 points), with a calculated window radius of r=200. When the algorithm processes the real-time sequence at i=500 (potentially corresponding to the mid-startup phase), it will not attempt to match points in the template where j=1 (early startup phase) or j=800 (steady-state operation). Instead, it will only consider points in the template where j ranges from 500-200=300 to 500+200=700. This forces the algorithm to find the best match among features near the "mid-startup phase," thereby eliminating a large amount of meaningless computation and ensuring the physical plausibility of the matching.

[0067] Please see Figure 3 As shown, the multi-feature weighted distance is defined as follows:

[0068]

[0069] In the formula, d ω (i,j) is the weighted composite distance of similarity between the i-th point in the real-time monitoring sequence and the j-th point in the reference template sequence, I i To monitor the effective current value at the i-th point in the sequence in real time, I j K is the effective value of the current at the j-th point in the reference template sequence. i To monitor the VI trajectory curvature at the i-th point in the sequence in real time, K j For the curvature of the VI trajectory at the j-th point in the reference template sequence, Let be the rate of change of current at the i-th point in the real-time sequence. Let be the rate of change of current at the j-th point in the reference template sequence.

[0070] In the improved DTW algorithm, d ω (i,j) is a weighted composite distance used to measure the similarity between the i-th point in the real-time sequence and the j-th point in the reference template. Unit: dimensionless scalar value. The smaller the value, the more similar the features of the two points. When d... ω If (i,j) is less than the threshold, then the two are successfully matched; otherwise, the match fails.

[0071] By employing a fast hierarchical strategy, the computational complexity is reduced from O(n^2). 2 The value decreases to O(n).

[0072] It also includes a transient process identification step, utilizing the device startup time t start and equipment downtime t end The transient time period of the k-th device is marked as Δt. k =[t start -0.5s, t end +1s]. The time extension (-0.5s, +1.0s) is set based on:

[0073] 1. (0.5s before startup) Integrity assurance: Used to capture the characteristics of the startup preparation stage, such as the movement of electrical contacts and the pre-power-on of control circuits, and to provide a steady-state reference for transient analysis, ensuring that the extraction of transient features is not affected by background interference.

[0074] 2. (1.0s after shutdown) Attenuation process capture: Used to fully record the current attenuation trajectory of the physical processes such as the dissipation of electromagnetic field energy and mechanical inertial rotation inside the equipment after the power is disconnected, to ensure the integrity of the transient energy consumption integral.

[0075] 3. Engineering applicability: This time parameter is based on the statistical analysis of the typical transient process durations of mainstream household appliances (air conditioners, refrigerators, motors) and industrial equipment, and can cover the transient processes of more than 95% of equipment, while avoiding the introduction of irrelevant steady-state data by using excessively long windows.

[0076] 4. Adaptive margin: The 0.5 seconds and 1.0 seconds mentioned are typical values. The system can be fine-tuned according to the calibration knowledge base of specific equipment types (e.g., the extension before starting a large motor can be increased to 0.8 seconds).

[0077] 3. Loss Compensation Calculation

[0078] Transient process energy compensation formula:

[0079]

[0080] In the formula, ΔE k For the transient compensation power of the k-th device, Let k be the startup time of the kth device. Let k be the downtime of the device. Let α be the rate of change of current for the k-th device. k Let be the compensation coefficient for the k-th device.

[0081] Compensation coefficient α k These are specific parameters for the k-th electrical device, obtained through laboratory calibration. The calibration method is as follows: under isolated power supply conditions, measure the actual total power consumption and steady-state power consumption of the device over a complete operating cycle; the difference between the two is the actual transient loss E. transient,true ; synchronously calculate its transient process through high-frequency sampling. Value as base loss E base The ratio of the two This is the compensation coefficient for this type of equipment, and it is pre-stored in the equipment feature library. This coefficient comprehensively characterizes the conversion losses of multiple energy sources, including mechanical energy and electromagnetic field energy, during the transient process of the equipment, in addition to Joule heating.

[0082] α kIt is a "calibrator" that connects the theoretical model of this invention with the real losses in the physical world. Its introduction enables the calculation model based on the current derivative to approximate the complex real energy consumption, which is the key to reducing the transient loss error from ±15% to ±3.2%.

[0083] α k It is a correction factor greater than 1. Its physical meaning is: to accurately measure the true total electrical energy consumed by the k-th device during a transient process, an additional amplification factor needs to be multiplied by the basic energy value calculated through the integral of the rate of change of current. In simpler terms: if the energy value calculated through the formula... The calculated value is considered as the "direct electrical and thermal consumption" of transient losses, then α k It is the "total loss amplification factor" used to compensate for other forms of loss (such as mechanical energy and electromagnetic energy conversion losses) that the formula has not fully captured.

[0084] Because of the core compensation items This primarily and accurately reflects the additional copper losses (Joule heating) caused by current surges. However, as mentioned earlier, the total transient losses also include:

[0085] ■ Mechanical acceleration loss (increase in motor rotor kinetic energy)

[0086] ■ Electromagnetic field establishment / discharge loss

[0087] ■ Switching device losses

[0088] These losses are not related to They exhibit a simple linear proportional relationship, but experiments show that they are strongly correlated with current surge losses. Therefore, α k As a comprehensive empirical coefficient, the measurable "integral of the square of the rate of change of current" is calibrated to be close to the true value of "the true total loss of the transient process", which is difficult to measure directly.

[0089] Please see Figure 5 As shown, α k Method for determining α k These are not theoretically derived values, but empirical data obtained through a rigorous "experimental calibration method." The calibration process is as follows:

[0090] 1. Isolation test: In a laboratory environment, power is supplied to a single target device (such as a certain model of air conditioner), and a high-precision standard energy meter is connected to the front end.

[0091] 2. Control and Measurement:

[0092] Make the device repeatedly cycle through "start-run-stop".

[0093] The high-frequency sampling unit of this invention is used to synchronously acquire the total current waveform I(t).

[0094] A standard electricity meter records the actual total power consumption E of the equipment over a complete cycle. total,measured .

[0095] 3. Calculation and Calibration:

[0096] The algorithm identifies the steady-state and transient periods of the device.

[0097] Calculate the energy consumption E during the steady-state period steady .

[0098] Subtracting the steady-state energy consumption from the total energy consumption yields the "measured transient total loss" E. transient,true =e total,measured -E steady .

[0099] Simultaneously, the collected I(t) is used to calculate the transient time period. Value, denoted as E base .

[0100] The compensation coefficient for this device is:

[0101]

[0102] Warehousing: Store this model of equipment α k The value and its characteristic fingerprint are stored together in the "Device Characteristic Library".

[0103] α of typical equipment k value:

[0104]

[0105] The compensation coefficient α of the kth device k Dynamic adjustment based on device type includes the following steps:

[0106] Step S1: Initial Matching: After identifying the type of the kth device, retrieve the initial value α of the calibration compensation coefficient corresponding to that device type from the pre-stored device feature library. preset ;

[0107] Step S2: Online Learning and Adaptive Correction: During the subsequent operation of the device, the actual energy consumption during its start-up and shutdown processes is continuously monitored. Based on multiple monitoring data, the compensation coefficient is adjusted using the following iterative algorithm:

[0108] Step S21: Calculate using the current coefficients Predicted transient loss E pred Compared with the actual back-calculated loss E actual The relative deviation δ between them;

[0109] Step S22: If the absolute value of the relative deviation average value If the threshold is exceeded, the compensation coefficient is adjusted according to the following formula.

[0110]

[0111] In the formula, μ is the learning rate. This represents the result of the t-th adjustment of the compensation coefficient. The result of the (t+1)th adjustment of the compensation coefficient is given, and δ is the predicted transient loss E. pred Compared with the actual back-calculated loss E actual The relative deviation between them;

[0112] Step S23: Update the It is written back to the device's feature record for subsequent calculations.

[0113] In this way, the compensation coefficient can achieve personalized initial matching based on equipment type, and can also be adaptively optimized according to the actual operating status of the equipment, thereby ensuring the long-term accuracy of transient loss compensation under different aging levels, load conditions and power grid environments.

[0114] The preset values ​​are obtained through calibration in an isolated laboratory. For example:

[0115] Brand A, Model X, 1.5 HP fixed-frequency air conditioner → α preset =1.12;

[0116] Brand B, 50kW three-phase asynchronous motor → α preset =1.09;

[0117] General purpose LED bulb → α preset =1.03.

[0118] Baseload energy consumption E base Real-time estimation is achieved through Kalman filtering.

[0119] 4. Daily electricity consumption calculation

[0120] Total power consumption calculation formula:

[0121]

[0122] In the formula, E base For base charge, ΔE k For the transient compensation power of the k-th device, P k (t) represents the power of the k-th device at time t.

[0123] Please see Figure 4 As shown, PCA feature decoupling is enabled under abnormal operating conditions:

[0124] C = F T ·W

[0125] Here, C is a p×n principal component score matrix, representing the decoupled output. Each row (one principal component) represents the "activation intensity" or characteristic of a potential, independent load source (such as an air conditioner or a motor) over time. Since the principal components are orthogonal to each other, the signals in each row of C are decoupled and independent.

[0126] F is the original characteristic matrix of n×m, F=[f1,f2,...,f m ] represents the input data, which is a mixed observation data consisting of m load characteristics (such as RMS current, harmonics, VI curvature, etc.) collected at n consecutive sampling times. Each column is a time series of one feature, and each row is a snapshot of all features at one time. F T Let F be the transpose of F.

[0127] W is an m×p load matrix (or projection matrix), W=[w1,w2,...,w p Its column vector w i The i-th principal component direction represents a new, mutually orthogonal coordinate axis in the original m-dimensional feature space. It is determined by the principal component contribution: 1. Contribution calculation: The proportion of the original data variance explained by each principal component is its contribution. The larger the contribution, the more original information the component carries. 2. Selection criteria: Typically, the top p principal components with a cumulative contribution exceeding 85%-95% are selected to form W (p≤m) to achieve dimensionality reduction and noise reduction.

[0128] Technical effects of the present invention:

[0129] 1. Precise compensation for transient losses

[0130] By quantifying the energy loss during transient processes such as equipment start-up, shutdown, and state switching using a quadratic integral compensation model, the transient error of ±15% in the traditional method is compressed to within ±3.2%, significantly improving the accuracy of power consumption calculation.

[0131] In this invention, the term "secondary" has a dual meaning:

[0132] 1. Mathematical quadratic: The core integral term of the model is the square of the rate of change of current (dI / dt). 2 .

[0133] 2. Second-order process in physics: This model describes the quadratic dependence from "current change" to "heat loss" (Joule's law P=I). 2 R), which is essentially a second-order energy conversion process.

[0134] II. Core Components and Derivation of the Model

[0135] The model aims to calculate the transient process ΔT of the k-th device. k Total loss ΔE k Its complete expression is:

[0136]

[0137] This model is derived and constructed through the following logical steps:

[0138] Step 1: Locate the core physics source (first-order integration relationship)

[0139] The highest proportion and easiest to quantify transient losses is "additional Joule heat loss caused by current surges".

[0140] According to Joule's law, the instantaneous heat power across the resistor R is P(t) = I(t). 2 R.

[0141] Therefore, the theoretical Joule heat loss is the integral of power:

[0142]

[0143] This integral is the model's first integral (integration with respect to time).

[0144] Step 2: Establish the connection with observable quantities (introduce differentiation)

[0145] During the transient process, the current I(t) changes drastically. Direct, high-precision measurement of I(t) is required. 2 The integral is greatly affected by noise.

[0146] For inductive loads such as motors and compressors, the rate of change of transient current dI / dt is highly correlated with its amplitude I, and dI / dt can more sharply capture the initiation of surge.

[0147] Through circuit analysis (e.g., the simplest RL circuit model ignoring back electromotive force), it can be deduced that at the instant of startup, the current rises approximately linearly: I(t)≈(dI / dt)·t.

[0148] Substituting into Joule's law formula, we get:

[0149]

[0150] At this point, the power P(t) is proportional to the square of the rate of change of current (dI / dt)². This is a direct representation of the "quadratic" nature of the equation.

[0151] Step 3: Construct a computable surrogate model (quadratic integral relation)

[0152] Integrating the above power expression, we obtain the energy:

[0153]

[0154] This is a more complex integral. To simplify and enhance its universality, the model underwent a key engineering simplification and innovation:

[0155] use It serves as a "core proxy variable" for characterizing the intensity of transient Joule heat loss.

[0156] This simplification is based on the following facts:

[0157] 1. (dI / dt) 2 A sharp peak appears at the beginning of the transient, and its integral value is naturally strongly correlated with the surge intensity.

[0158] 2. This form of calculation is stable and has a certain degree of noise suppression (differentiation itself is a high-pass filter).

[0159] 3. It avoids direct reliance on the difficult-to-obtain load resistance R and complex t 2 item.

[0160] Thus, we obtain the core integral term of the model: This can be understood as accumulating a continuous measure of the "drastic change in current" (second integration).

[0161] Step 4: Introduce compensation coefficients to complete the model (hybrid modeling)

[0162] Core Integral Items It primarily reflects Joule heating. However, as mentioned earlier, the total transient loss also includes mechanical energy, electromagnetic energy, and other components.

[0163] Therefore, a device-specific compensation coefficient α is introduced. k This involves "amplifying" or "calibrating" the core integral term to approximate the true total loss.

[0164] The final model was established as follows:

[0165]

[0166] Compensation coefficient adaptive calibration: The α coefficient is dynamically adjusted based on the equipment type (such as air conditioner, motor) to avoid human experience intervention (such as α = 1.12 for air conditioner, α = 1.08 for motor).

[0167] 2. High-precision decoupling of composite loads

[0168] By employing principal component analysis (PCA) and multi-feature weighted DTW algorithm, load feature separation is achieved under unsupervised conditions for parallel operation of multiple devices, improving the accuracy of composite scene identification from 78% to 93.6%.

[0169] It supports independent energy consumption accounting for complex scenarios such as industrial motor groups and residential charging piles.

[0170] 3. Computational efficiency is greatly improved.

[0171] The improved window constraint DTW algorithm reduces the computational complexity from O(n) 2 The time complexity is reduced to O(n), and the matching delay is shortened from 5.2 seconds to 0.8 seconds, meeting the real-time requirements.

[0172] Hierarchical recursive strategy: By predicting coarse-grained paths, redundant calculations are reduced by up to 70%.

[0173] Practical application advantages

[0174] 1. Unsupervised learning reduces implementation costs.

[0175] No manual annotation of device feature libraries is required; load fingerprints are automatically constructed through PCA and dynamic clustering, reducing annotation costs by 100%.

[0176] Supports automatic updates of the feature library after adding or replacing devices (e.g., automatically registering features for newly purchased air conditioners).

[0177] 2. High-frequency sampling enhances scene adaptability

[0178] Employing ≥2kHz high-frequency sampling technology, it captures transient features such as current change rate (dI / dt) and VI trajectory curvature, making it suitable for high dynamic load scenarios such as air conditioners and electric vehicle charging piles.

[0179] It supports complex power grid environments such as voltage fluctuations (±10%) and harmonic distortion (THD≤8%).

[0180] 3. Significant economic benefits

[0181] Improved line loss location accuracy: By calculating the power consumption of each item, high-loss lines can be accurately located, reducing the line loss rate by 0.3-0.8 percentage points (pilot data).

[0182] User energy efficiency optimization: Provides equipment-level power consumption reports to guide users in adjusting their power consumption behavior. Actual measurements for residential users show an average power saving rate of 7.2%.

[0183] Advantages compared to traditional methods

[0184]

[0185] Example 1 (Residential User)

[0186] 1. Sampling rate set to 4.8kHz, detecting the transient response during air conditioner startup.

[0187] 2. The compressor load is determined by the VI curvature κ > 0.15;

[0188] 3. Calculate steady-state energy consumption

[0189] 4. Compensation for transient losses ΔE = 1.12 × 0.8 kWh = 0.896 kWh;

[0190] 5. Total power E day =13.5+0.896+2.1(base load)=16.496kWh.

[0191] Example 2 (Industrial Motor Group)

[0192] 1. Monitor the harmonic characteristic matrix F = [f3rd, f5th, f7th] of 3 motors;

[0193] 2. Separate the superimposed loads using PCA, with weights W = [0.7, 0.2, 0.1];

[0194] 3. Calculate the energy consumption contribution of each motor: E1 = 42kWh, E2 = 37kWh, E3 = 15kWh;

[0195] 4. Total electricity consumption Eday = 42 + 37 + 15 = 94 kWh.

[0196] Technical effects:

[0197]

[0198]

Claims

1. A method for determining daily electricity consumption based on dynamic load identification, characterized in that, Includes the following steps: Step 1: High-frequency feature acquisition: Acquire the current waveform I(t) and voltage waveform V(t) at a sampling rate α, where α ≥ 2kHz; extract the feature vector F using the obtained current waveform I(t) and voltage waveform V(t). k The feature vector F k Including harmonic energy distribution, VI trajectory curvature, and current change rate. Step 2, Load the feature library: Call the reference feature matrix F ref =[F1,F2…F m The window-constrained dynamic time warping (DTW) algorithm is used to analyze the eigenvector F. k and reference characteristic matrix F ref The matching process is performed to obtain the matching results, which include the device ID and the device startup time t. start and equipment downtime t end , where the window constraint is: path offset radius: r = [0.2 × max(n, m)], where n is the length of the real-time monitoring sequence and m is the length of the reference template sequence; Path offset range: |ij|≤r, where i is the i-th sampling point in the real-time monitoring sequence and j is the j-th sampling point in the reference template sequence; Multi-feature weighting: The distance is calculated using the following formula: In the formula, d ω (i,j) is the weighted composite distance of similarity between the i-th point in the real-time monitoring sequence and the j-th point in the reference template sequence, I i To monitor the effective current value at the i-th point in the sequence in real time, I j K is the effective value of the current at the j-th point in the reference template sequence. i To monitor the VI trajectory curvature at the i-th point in the sequence in real time, K j For the curvature of the VI trajectory at the j-th point in the reference template sequence, Let be the rate of change of current at the i-th point in the real-time sequence. The current change rate at the j-th point in the reference template sequence; Step 3: Determine the daily electricity consumption E using the following formula. day : In the formula, E base For base charge, ΔE k For the transient compensation power of the k-th device, Let k be the startup time of the kth device. Let k be the downtime of the device. Let α be the rate of change of current for the k-th device. k Let P be the compensation coefficient for the k-th device. k (t) represents the power of the k-th device at time t.

2. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, Base charge E base Real-time estimation via Kalman filtering: AND base =∑P base .Δt In the formula, P base Δt represents the base load power, and Δt represents the base load operating time.

3. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, The curvature of the VI trajectory in step 1 is expressed by the following formula: In the formula, κ is the curvature of the VI trajectory, V is the voltage value, and I is the current value. The first derivative of voltage with respect to time. This is the second derivative of voltage with respect to time. This is the first derivative of the current with respect to time. It is the second derivative of the current with respect to time.

4. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, Step 2 also includes a transient process identification step, utilizing the device startup time t. start and equipment downtime t end The transient time period of the k-th device is marked as Δt. k =[t start -0.5s, t end +1s].

5. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, When matching fails in step 2, principal component analysis (PCA) is used to decouple features and separate composite load features. C=F T ·W In the formula, C is the p×n principal component score matrix, W is the m×p loading matrix, F is the n×m original feature matrix, and F T Let F be the transpose of F.

6. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, The compensation coefficient α of the kth device k Dynamic adjustment based on device type includes the following steps: Step S1: Initial Matching: After identifying the type of the kth device, retrieve the initial value α of the calibration compensation coefficient corresponding to that device type from the pre-stored device feature library. preset ; Step S2: Online Learning and Adaptive Correction: During the subsequent operation of the device, the actual energy consumption during its start-up and shutdown processes is continuously monitored. Based on multiple monitoring data, the compensation coefficient is adjusted using the following iterative algorithm: Step S21: Calculate using the current coefficients Predicted transient loss E pred Compared with the actual back-calculated loss E actual The relative deviation δ between them; Step S22: If the absolute value of the relative deviation average value If the threshold is exceeded, the compensation coefficient is adjusted according to the following formula. In the formula, μ is the learning rate. This represents the result of the t-th adjustment of the compensation coefficient. The result of the (t+1)th adjustment of the compensation coefficient is given, and δ is the predicted transient loss E. pred Compared with the actual back-calculated loss E actual The relative deviation between them; Step S23: Update the It is written back to the device's feature record for subsequent calculations.

7. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, The compensation coefficient for the air conditioner is 1.12, and the compensation system for the motor is 1.

05.

8. The method for determining daily electricity consumption based on dynamic load identification as described in claim 1, characterized in that, Step 3 also includes an anomaly detection step, which calculates the daily electricity consumption E obtained in step 3. day and total meter reading E meter The deviation rate δ, when the deviation rate δ≤5%, then the daily electricity consumption E is output. day Otherwise, repeat steps 2 and 3.

Images